Each solid blue bar is one period's real sales volume: a trend component, a seasonal cycle and random market noise, generated fresh every tick and scrolling toward the viewer as time passes. The glowing node above the "now" plane is the AI forecasting model — it never sees the future bars (translucent, ahead of the plane); it only re-estimates its internal state from bars that have already arrived, then projects forward. It uses Holt's linear (double) exponential smoothing: a running level and trend that update every period from the newest actual value, weighted by the adaptivity slider (α). The forecast bars are the model's projection for the next N periods, with a translucent envelope showing its uncertainty band, which widens the further out it looks.
level_t = α·actual_t + (1-α)·(level_{t-1} + trend_{t-1})
trend_t = β·(level_t - level_{t-1}) + (1-β)·trend_{t-1}
forecast_{t+h} = level_t + h·trend_t (± growing uncertainty band)
- Trend strength — the long-run slope of real demand; negative means a shrinking market.
- Seasonality — amplitude of a repeating demand cycle (e.g. quarterly buying patterns) layered on top of the trend.
- Market noise — random demand shocks the model cannot predict, the honest floor on any forecast's accuracy.
- Adaptivity (α) — how fast the model's internal level/trend estimate reacts to new data: low α smooths past noise but reacts slowly to real shifts; high α tracks change fast but chases noise.
- Forecast horizon — how many periods ahead the model projects; error compounds with horizon, visible as the widening translucent band.
- Demand shock — injects a sudden one-off jump/drop, the kind of real-world disruption (a competitor launch, a stockout, a viral trend) that tests how fast the model recovers.
Real-world relevance: this is the same trade-off every production sales-forecasting system faces — smoothing enough to ignore noise, but adaptive enough to catch genuine shifts in demand before they become costly stockouts or overstock.